acceptodds
Under review as a conference paper at ICLR 2027

Evidence Is Layered: Expert-Guided Queries for Generalizable Image Forgery Detection

Abstract

Pretrained vision transformers (ViTs) provide transferable representations for generalizable image forgery detection, yet how forensic evidence is distributed across their layers and how to retrieve it effectively remain poorly understood. We investigate these questions with LayerProbe, a controlled study of 14 frozen ViT backbones spanning standalone and MLLM vision encoders, using layer-wise linear probing and greedy layer fusion. The results reveal that the best readout depth varies across backbones, while selected layers with lower single-layer scores can still complement the strongest layer. Motivated by these findings, we propose ForenLens, a frozen-backbone detector that uses queries to retrieve complementary forensic evidence from a compact, backbone-specific layer set selected by LayerProbe. ForenLens uses a shared cross-attention reader to extract layer-specific responses before combining them through query-adaptive routing, preserving layer identity and learning where each query reads for each image. Complementing this adaptive layer routing, training-only forensic guidance shapes what expert queries retrieve by supervising their spatial attention. On the OpenMMSec out-of-distribution split and seven external benchmarks, ForenLens achieves state-of-the-art performance with 93.2% mean AP and 87.1% mean accuracy, while remaining robust to post-processing. Our code will be available upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Forgery Detection, AI-generated Image Detection, Vision Transformers, Out-Of-Distribution Generalization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_aBB2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how to extract forgery-detection cues from frozen pretrained visual encoders. LayerProbe uses layer-wise linear probing and greedy fusion across 14 ViT backbones to identify useful, complementary layers. The results show that the best readout depth varies across backbones and that individually weaker layers can improve a combined readout. Building on these findings, ForenLens uses shared cross-attention and image-dependent layer routing to retrieve evidence from a fixed set of selected layers. Four forensic priors guide query attention during training, while a free query is intended to capture information beyond these priors.

Critical strengths

I find the idea of extracting forgery-detection cues already present in pretrained visual encoders convincing. Investigating where these cues can be read and how to retrieve them effectively is a worthwhile direction.

Critical weaknesses

My main concern is how well the selected layers, encoder-specific readout, and predefined expert priors remain useful when conditions change. The paper aims at generalization, but its implementation still depends on these choices. The reported OOD results support the tested configuration; they do not fully address how these choices hold up when the source data, encoder, or generation mechanism changes.

  1. Layer selection depends on the available forgery data, but the paper does not compare results when the source data used for selection changes. For example, with CLIP fixed, would LayerProbe select the same layers when run separately on AIGC, Deepfake, and image manipulation data? How would those layer sets perform on the same unseen test set? Table 6 tests OOD transfer after selection on OpenMMSec, but does not answer how much the selected layers and their benefits depend on the forgery types in the source data.

  2. Layer selection is the foundation of the method, and it is tied to the visual encoder. The paper evaluates 14 encoders, but performs probing, selection, and readout training separately for each. This demonstrates that the same design can be trained with different encoders; it does not demonstrate direct reuse of the selected layers or trained readout across encoders. Following the proposed procedure, changing the encoder still requires these adaptation steps.

  3. Once selection is complete, dynamic routing can only reweight the selected layers; it cannot add other layers as readout sources. Its scope is therefore determined by the fixed selection. The default of four layers also lacks sufficient justification. The reported comparisons cover one to four layers, with no results for five or more. In Table 20, SigLIP2 achieves 93.6% mean AP with two layers, compared with 93.2% with four. These results show that multiple layers help, but do not explain why four should be the default across encoders.

  4. The coverage of the expert priors for fully generated images needs clarification. Images generated without local splicing or editing do not necessarily contain tampering boundaries or locally inconsistent compression histories. For an example such as GPT-Image2.5, which I am concerned about, the paper reports no dedicated evaluation. Not all four experts target local manipulation: SD-VAE reconstruction error is also used to analyze generated images, but it relies on a fixed autoencoder. The results do not fully establish when these predefined priors remain informative across different generation mechanisms. This is a concern about their coverage, rather than evidence that the detector already fails on that particular generator.

Questions for authors

No questions. I would like the authors to reconsider the paper’s contributions and how they are presented.

7 comments
AuthorsOpenReview ↗

Official Comment by Authors Part1

We thank the reviewer for the constructive feedback and address the four concerns below.

W1 (Selection source sensitivity)

We ran LayerProbe separately on each single source domain in OpenMMSec (AIGC, Deepfake, IMDL, and Document), following the settings of Table 6.

BackboneSelection SourceSelected BlocksOOD S1S_1 AUCOOD S4S_4 AUCΔ\Delta AUC
CLIPAIGC-only10, 11, 23, 177.5685.03+7.47
CLIPDeepfake-only11, 21, 9, 1278.3385.26+6.93
CLIPIMDL-only8, 21, 11, 774.8884.34+9.46
CLIPDocument-only8, 23, 7, 974.8884.32+9.44
CLIPAll four domains9, 22, 14, 776.7186.07+9.36
SigLIP2AIGC-only27, 19, 16, 2582.7386.65+3.92
SigLIP2Deepfake-only14, 9, 24, 1180.2387.29+7.06
SigLIP2IMDL-only6, 17, 12, 2774.5288.03+13.51
SigLIP2Document-only9, 13, 26, 577.6887.78+10.10
SigLIP2All four domains11, 26, 7, 1479.0088.41+9.41

Across all settings, LayerProbe consistently improves over the single-layer baseline on OOD test split. Because adjacent ViT blocks share similar features (see xWc2 - W2&Q2), the exact block indices naturally vary by probe domain, yet almost all settings select a mix of early/middle and deep blocks. The only exception is SigLIP2 probed on AIGC-only, which focuses on middle-to-deep blocks because its final block alone is already strong on AIGC. Consequently, it gives the smallest gain, whereas probing across all four domains achieves the highest overall OOD AUC among the source-selection settings shown.

W2 (Backbone adaptation)

As stated in the main text (Sections 1 and 4.1), ForenLens is a general architectural design across diverse ViT backbones, rather than a fixed module reused across encoders without training. Similar to methods such as LoRA (ICLR'22 [1]), which fit lightweight parameters to each frozen model, adapting the layer selection and readout to each frozen ViT is standard for a general framework.

W3 (Choice of K=4K=4)

We extended the LayerProbe and 8-Group experiments up to K=10K=10.

Model / MetricK=1K=1K=2K=2K=3K=3K=4K=4K=5K=5K=6K=6K=7K=7K=8K=8K=9K=9K=10K=10
LayerProbe OOD AUC (CLIP)76.783.385.886.186.386.786.686.586.586.5
LayerProbe OOD AUC (SigLIP2)79.086.988.088.488.588.688.788.888.988.9
ForenLens 8-Group AP (CLIP)79.580.681.083.582.982.481.781.082.082.3
ForenLens 8-Group AP (SigLIP2)81.893.693.493.293.293.792.192.593.492.0

In LayerProbe, multi-layer combinations (K≥2K\ge 2) consistently outperform single-layer baselines. Beyond K=4K=4, the marginal gains from further increasing KK diminish, and larger KK does not consistently improve ForenLens.

KK is a fixed hyperparameter, and incremental readout schemes that support a dynamic KK are beyond the scope of our current design. Across backbones, the optimal KK varies and balances accuracy, routing difficulty, and compute cost. We use K=4K=4 as a shared, compact layer budget across all 14 backbones, without tuning KK separately for each backbone (which is also why Table 1 reports K=4K=4 for SigLIP2 rather than the higher-scoring K=2K=2).

AuthorsOpenReview ↗

Official Comment by Authors Part2

W4 (Fully generated images)

  1. Fully generated images also exhibit artifacts in local regions, as studied in AutoGAN (WIFS'19 [2]), CNNDetection (CVPR'20 [3]), and NPR (CVPR'24 [4]). Likewise, reconstruction error from a fixed pretrained model remains informative across unseen generation mechanisms on fully generated images, as demonstrated in DIRE (ICCV'23 [5]), AEROBLADE (CVPR'24 [6]), and DRCT (ICML'24 [7]).
  2. The benchmark AIGI-Now (AIGI-P and AIGI-S) already evaluates 9 modern generative models, including GPT-4o, Nano Banana, FLUX Pro, FLUX Dev, FLUX Kontext, FLUX Krea, Jimeng, Kling, and MiniMax. The fake samples from these generators in AIGI-P and AIGI-S are fully generated images. On AIGI-P and AIGI-S, ForenLens with SigLIP2 achieves 95.6% and 89.7% AP, respectively. We will add the version and complete generator list to the revised manuscript.

References

[1] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-Rank Adaptation of Large Language Models. In International Conference on Learning Representations, 2022.

[2] Xu Zhang, Svebor Karaman, and Shih-Fu Chang. Detecting and Simulating Artifacts in GAN Fake Images. In IEEE International Workshop on Information Forensics and Security, pages 1–6, 2019.

[3] Sheng-Yu Wang, Oliver Wang, Richard Zhang, Andrew Owens, and Alexei A. Efros. CNN-Generated Images Are Surprisingly Easy to Spot... for Now. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8695–8704, 2020.

[4] Chuangchuang Tan, Huan Liu, Yao Zhao, Shikui Wei, Guanghua Gu, Ping Liu, and Yunchao Wei. Rethinking the Up-Sampling Operations in CNN-based Generative Network for Generalizable Deepfake Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28130–28139, 2024.

[5] Zhendong Wang, Jianmin Bao, Wengang Zhou, Weilun Wang, Hezhen Hu, Hong Chen, and Houqiang Li. DIRE for Diffusion-Generated Image Detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22445–22455, 2023.

[6] Jonas Ricker, Denis Lukovnikov, and Asja Fischer. AEROBLADE: Training-Free Detection of Latent Diffusion Images Using Autoencoder Reconstruction Error. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9130–9140, 2024.

[7] Baoying Chen, Jishen Zeng, Jianquan Yang, and Rui Yang. DRCT: Diffusion Reconstruction Contrastive Training towards Universal Detection of Diffusion Generated Images. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 7621–7639, 2024.

AuthorsOpenReview ↗

We look forward to your responses

Dear Reviewer aBB2,

We have provided point-wise clarification and experiments (upon your request) for all your concerns. We look forward to your reply and hope that our responses could address your concerns. Your time and effort is very much appreciated.

Best Regards

Reviewer_xWc2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how forensic evidence is distributed across layers of pretrained Vision Transformers (ViTs) and how cross-layer representations can improve generalization to unseen forgery types. The authors introduce LayerProbe, which performs layer-wise linear probing and greedy layer fusion across 14 frozen ViT backbones. The analysis reveals backbone-dependent optimal readout depths and shows that individually weaker layers can provide complementary discriminative information.

Building on these findings, the authors propose ForenLens, a frozen-backbone detector that uses learnable queries and query-adaptive layer routing to retrieve and integrate information from selected layers. Training-only forensic priors supervise the spatial attention of expert-guided queries, while a free query with orthogonality regularization is designed to capture information beyond predefined expert priors. Experiments on OpenMMSec OOD and seven external, primarily AIGC-oriented benchmark groups report competitive detection performance, supported by ablations, matched-backbone comparisons, and robustness evaluations.

Critical strengths
  1. The paper provides a systematic empirical characterization of layer-wise forensic discriminability and cross-layer complementarity across 14 pretrained ViT backbones, including standalone and MLLM vision encoders.

  2. ForenLens presents a coherent integration of cross-layer query-based readout, adaptive routing, and training-only forensic guidance. Its frozen-backbone design achieves competitive performance with reduced backbone-training costs.

  3. The paper evaluates the complete framework across multiple backbones and benchmarks, with ablations, matched-backbone comparisons, and robustness analyses providing useful empirical support for its overall effectiveness.

Critical weaknesses
  1. The evaluation protocol may overstate generalization across diverse forgery domains. Although the paper targets generalizable image forgery detection and trains on OpenMMSec, which covers four forensic domains, the overall score equally weights OpenMMSec and seven external AIGC benchmark groups. Consequently, these AIGC-oriented benchmarks account for 87.5% of the aggregate score, potentially overstating the evidence for generalization across diverse forgery domains. For example, in Table 15, ForenLens (SigLIP2) achieves 84.7% ACC on OpenMMSec, slightly below SICA's 85.4%, yet improves the eight-group mean ACC from 77.1% (SICA) to 87.1% (ForenLens).

  2. LayerProbe demonstrates discriminative gains but does not rule out semantic or dataset-specific cues as alternative explanations. This gap directly undermines the empirical premise motivating ForenLens.

  3. The layer-selection evaluation relies on a different readout mechanism from ForenLens, leaving it unclear whether greedy layer selection remains advantageous within the final framework. Furthermore, the comparison between query-based readout and pooling does not fully control for model capacity and auxiliary supervision.

  4. The forensic reliability of expert guidance and the claimed query specialization remain unverified. Guidance signals may capture benign structures, while similar routing patterns raise doubts about distinct query roles. Existing ablations do not establish the forensic specificity or functional specialization of these queries.

I appreciate the additional experiments, but my main concerns remain unresolved.

The revised aggregation provides no new evidence of non-AIGC generalization. Correlation analyses do not exclude semantic or dataset-specific cues, and decision weights do not establish artifact-specific query responses. Therefore, I maintain my socre.

Questions for authors
  1. The paper should more clearly define its focus on generalizable image forgery detection. If the goal is generalization across diverse forensic domains, the evaluation should reduce the disproportionate weight of AIGC benchmarks or incorporate additional test sets covering other domains (IMDL, Deepfake, and Document) to provide a more balanced assessment.
  2. Could the authors provide controlled experiments to establish that cross-layer gains arise from complementary forensic artifacts rather than semantic or dataset-specific cues?
  3. Regarding Weakness 3, could the authors compare greedy selection with alternative strategies within ForenLens and isolate the benefits of query-based readout through capacity-matched comparisons?
  4. Could the authors validate the forensic reliability of expert guidance and demonstrate that expert-guided queries respond to their intended artifacts while the free query captures complementary evidence?
4 comments
AuthorsOpenReview ↗

Official Comment by Authors Part1

We thank the reviewer for the constructive comments and address each point below.

W1&Q1 (Domain-balanced evaluation)

OpenMMSec covers 72 unseen forgery subtypes across four domains. We evaluate our method under two domain-balanced protocols where Deepfake, IMDL, Document, and AIGC each carry an equal 25% weight.

  1. On OpenMMSec OOD test sets (Appendix Tables 10–13), ForenLens (SigLIP2) outperforms SICA in AP, ROC-AUC, and F1.
  2. We first average the OpenMMSec AIGC domain and seven external benchmark groups into a single Unified-AIGC score (25%) and average it equally with Deepfake (25%), IMDL (25%), and Document (25%). As seen below, ForenLens (SigLIP2) outperforms SICA and all other baselines across all four metrics, including Accuracy.
MethodOpenMMSec (4-Domain Equal) AP / AUC / F1 / AccGlobal 4-Domain Balanced AP / AUC / F1 / Acc
RINE73.3 / 82.3 / 62.9 / 81.174.0 / 77.7 / 64.2 / 76.5
Effort73.1 / 84.1 / 62.8 / 80.170.4 / 74.3 / 59.4 / 74.0
PGC75.3 / 85.9 / 65.1 / 83.474.6 / 81.5 / 64.5 / 78.4
SICA79.0 / 87.5 / 69.7 / 85.478.6 / 84.4 / 69.5 / 81.4
ForenLens (CLIP)78.8 / 87.5 / 69.8 / 85.179.4 / 85.1 / 68.4 / 81.3
ForenLens (SigLIP2)81.8 / 89.0 / 71.3 / 84.782.8 / 88.2 / 72.8 / 82.9

W2&Q2 (Cross-layer complementarity)

  1. We analyze the class-conditional correlation between the greedily selected shallow (ℓa\ell_a) and deep (ℓb\ell_b) probes versus an adjacent-layer control on unseen OOD data, following Table 6. Here, b→ab\to a is the fraction of layer aa's errors corrected by layer bb, and Δ\Delta is the gain over the better single layer.
Backbone (Unseen OOD)Layer Pair (ℓa,ℓb\ell_a, \ell_b)Class-Cond. Pearson rr (Mean)Prediction DisagreementErrors Rescued (b→ab\to a / a→ba\to b)Single ℓa\ell_a / ℓb\ell_b AUCFused AUC (Δ\Delta)
CLIPAdjacent (9, 8)0.61222.3%35.1% / 38.4%76.7 / 74.977.2 (+0.5)
CLIPSelected (9, 22)0.25233.2%56.1% / 56.3%76.7 / 77.583.3 (+5.8)
SigLIP2Adjacent (11, 10)0.59822.9%39.0% / 40.8%79.0 / 77.981.1 (+2.1)
SigLIP2Selected (11, 26)0.35228.9%57.4% / 51.3%79.0 / 82.886.9 (+4.1)

As shown, adjacent layers have high class-conditional correlation and minor fusion gains, whereas the selected shallow and deep layers have lower class-conditional correlation and rescue over half of each other's errors.

  1. On unseen regional manipulation subsets, fusing the four selected layers (S4S_4 vs. S1S_1) on CLIP yields large OOD AUC gains: +16.4 points on Non-Generative IMDL (61.5% →\to 77.9%), +28.6 points on Face Editing (59.6% →\to 88.2%), +28.0 points on Face Swapping (57.6% →\to 85.6%), and +18.2 points on Face Reenactment (64.5% →\to 82.7%), with consistent gains ranging from 15.7 to 20.1 AUC points on SigLIP2. These show that cross-layer benefits transfer to unseen regional forgeries. Our evaluation covers 72 unseen forgery types and 51 external test sets, with cross-layer transfer also tested across 14 encoders (Tables 1 and 6, Appendix D.1).
AuthorsOpenReview ↗

Official Comment by Authors Part2

W3&Q3 (Layer selection and readout controls)

In addition to the LayerProbe comparisons against Rank-4 and Repeat-ℓ1\ell_1 in Appendix Table 6, we evaluate Rank-4 under the main experimental setting and compare query readout with and without expert supervision against capacity-matched pooling controls (Concat-Pooled + MLP-PM).

Variant (K=4K=4)Params (CLIP / SigLIP2)Expert SupervisionCLIP AP / AUCSigLIP2 AP / AUC
Concat-Pooled + Linear0.02M / 0.02M×\times79.5 / 81.588.9 / 89.6
Concat-Pooled + MLP-PM (capacity-matched)8.42M / 10.66M×\times81.0 / 81.090.5 / 90.3
ForenLens w/ 5 Free Queries (unguided)8.42M / 10.65M×\times80.1 / 78.790.9 / 91.2
ForenLens w/o L_guide8.42M / 10.65MLorth+Lrec\mathcal{L}_{\mathrm{orth}} + \mathcal{L}_{\mathrm{rec}}80.5 / 79.490.5 / 91.1
ForenLens (Rank-4, Full)8.42M / 10.65M✓\checkmark79.5 / 81.491.6 / 91.8
ForenLens (Greedy-4, Full)8.42M / 10.65M✓\checkmark83.5 / 83.893.2 / 93.3

As seen, Greedy-4 substantially outperforms Rank-4 on CLIP, where Rank-4 selects adjacent middle layers and misses deep complementary features. At matched parameter counts, unguided queries (5 Free Queries) already surpass Concat-Pooled + MLP-PM on SigLIP2, and adding expert supervision improves both backbones further.

W4&Q4 (Expert guidance and query specialization)

  1. That is correct. Guidance signals can indeed respond to benign structures in real images, which is why we use them as "guidance" rather than as direct classifiers. As shown in Appendix H (Figure 9, real image row), queries also attend to candidate regions in real images. They extract features from these candidate regions, and the experts teach them where the candidate regions are.
  2. These forensic experts use operators or principles already validated in forgery detection, including noise fingerprints (TruFor, CVPR'23 [1]), JPEG compression traces (Farid, TIFS'09 [2], CAT-Net, IJCV'22 [3]), Sobel-based boundary cues (MVSS-Net, ICCV'21 [4]), and reconstruction error (AEROBLADE, CVPR'24 [5]). Our contribution is not proposing these experts, but combining them as guidance.
  3. On the validation set (CLIP), we measure the spatial alignment between each query's attention map and the expert guidance maps. The four expert-guided queries align closely with their respective target maps (cosine similarity: Noiseprint++: 83.5%, Boundary: 87.8%, SD-VAE: 97.9%, JPEG Ghost: 98.5%), whereas the free query shows much lower similarity with all four expert maps (21.4%–34.3%).
  4. Regarding complementarity, Appendix G.4.2 (Table 24) shows that the decision readout assigns distinct mean weights across domains and real/fake samples. For instance, the Boundary query weight reaches 67.2% on Deepfake fake samples, but drops to 26.4% on AIGI-S and 29.2% on Chameleon. This shows that the readout adaptively combines query features across different scenarios.

References

[1] Fabrizio Guillaro, Davide Cozzolino, Avneesh Sud, Nicholas Dufour, and Luisa Verdoliva. TruFor: Leveraging All-Round Clues for Trustworthy Image Forgery Detection and Localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20606–20615, 2023.

[2] Hany Farid. Exposing Digital Forgeries From JPEG Ghosts. IEEE Transactions on Information Forensics and Security, 4(1):154–160, 2009.

[3] Myung-Joon Kwon, Seung-Hun Nam, In-Jae Yu, Heung-Kyu Lee, and Changick Kim. Learning JPEG Compression Artifacts for Image Manipulation Detection and Localization. International Journal of Computer Vision, 130(8):1875–1895, 2022.

[4] Xinru Chen, Chengbo Dong, Jiaqi Ji, Juan Cao, and Xirong Li. Image Manipulation Detection by Multi-View Multi-Scale Supervision. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14185–14193, 2021.

[5] Jonas Ricker, Denis Lukovnikov, and Asja Fischer. AEROBLADE: Training-Free Detection of Latent Diffusion Images Using Autoencoder Reconstruction Error. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9130–9140, 2024.

AuthorsOpenReview ↗

We look forward to your responses

Dear Reviewer xWc2,

We have provided point-wise clarification and experiments (upon your request) for all your concerns. We look forward to your reply and hope that our responses could address your concerns. Your time and effort is very much appreciated.

Best Regards

Thank you for the follow-up. Below are new external non-AIGC experiments and clarifications on the three remaining points:

  1. In Q1, you suggested that "the evaluation should reduce the disproportionate weight of AIGC benchmarks or incorporate additional test sets covering other domains (IMDL, Deepfake, and Document)." Our previous response adopted the reweighting option you proposed, reducing the AIGC weight from 87.5% to an equal 25% alongside Deepfake, IMDL, and Document. Under this setting, ForenLens outperformed SICA and all other baselines. We respectfully note that this reweighting directly followed the first option in Q1. As the follow-up now calls for new non-AIGC evidence rather than a rebalanced aggregation, we additionally implement the second option below.

    We evaluated the fixed OpenMMSec-trained Table 1 detectors, without retraining or layer reselection, on three external non-AIGC benchmarks disjoint from OpenMMSec: WildDeepfake [1] (Deepfake), DocTamper-CD [2, 3] (Document), and TFR-SceneText [3] (IMDL, non-generative copy-move/splicing). AP / Accuracy (%):

    MethodWildDeepfakeDocTamper-CDTFR-SceneTextMean
    CNNDetection59.6/58.543.0/37.549.2/48.950.6/48.3
    NPR66.5/59.450.5/50.451.6/49.956.2/53.2
    PGC70.5/64.942.4/42.154.4/53.055.8/53.3
    Effort73.9/66.345.0/43.655.0/52.258.0/54.0
    UnivFD71.7/65.047.0/47.856.7/52.558.5/55.1
    SICA74.4/65.544.1/47.867.5/61.862.0/58.4
    ForenLens (SigLIP2)79.6/65.944.3/47.478.2/68.867.3/60.7
    ForenLens (CLIP)79.3/70.163.4/55.972.6/60.971.8/62.3

    Both ForenLens variants outperform SICA and all compared baselines in mean AP and Accuracy, which we believe also addresses the new concern raised in your follow-up.

  2. Beyond correlation analyses, cross-layer gains are also supported by the controlled design of AIGI-Now (AIGI-P and AIGI-S). Both tracks start from real COCO test images and use the same nine generators. AIGI-P generates semantically faithful counterparts from the original COCO captions and matches their JPEG compression to the real images. This suppresses semantic and format shortcuts, so detection relies on low-level pixel artifacts. AIGI-S instead pairs real images with semantically implausible generations and applies heavy post-processing (downsampling, noise, blur, and recompression) to both classes. This wipes out low-level traces, so detection relies on high-level semantic cues.

    SigLIP2 Variant (Capacity-Matched)AIGI-P AP/AUCAIGI-S AP/AUC8-Group Mean AP/AUC
    Single-Layer (Top-1, 10.65M)92.7/92.157.5/52.581.8/82.6
    Concat-Pooled + MLP-PM (4 layers, 10.66M)88.6/86.484.3/79.190.5/90.3
    ForenLens (4 layers, Full, 10.65M)95.6/94.689.7/85.793.2/93.3

    On AIGI-P, ForenLens clearly improves over single-layer readout, whereas flat four-layer pooling (Concat-Pooled + MLP-PM) washes out localized pixel cues and drops below the single layer. On AIGI-S, single-layer readout falls near chance, while ForenLens leverages deeper features. ForenLens thus surpasses both baselines on both tracks. Together with our OOD gains on localized manipulations (W2&Q2, Point 2) and single-domain layer-selection transfer (aBB2 - W1), this provides direct controlled evidence beyond inter-layer correlations.

  3. Regarding W4&Q4, your comment on decision weights addresses Point 4 of our previous reply. Points 1–3 and our guidance ablations separately support expert guidance and query specialization. In Points 1–2, we clarified that heuristic operators also respond to benign structures. ForenLens therefore uses established forensic operators only as training-time spatial priors over candidate regions, not as test-time classifiers. In Point 3, validation results show that the four guided queries align closely with their target maps, while the orthogonal free query attends to distinct regions. Our controlled ablations (Tables 2–3 and e1Ua - W2&Q2) further show that shuffling guidance maps across images within each minibatch degrades performance even below removing spatial guidance altogether. Dropping any single query also hurts both backbones. These results confirm that aligned priors provide meaningful forensic supervision beyond generic regularization.

In light of these new external non-AIGC evaluations and clarifications, we kindly hope you might reconsider your assessment.

References

  • [1] B. Zi et al. WildDeepfake: A Challenging Real-World Dataset for Deepfake Detection. ACM MM, 2020.
  • [2] C. Qu et al. Towards Robust Tampered Text Detection in Document Image: New Dataset and New Solution. CVPR, 2023.
  • [3] C. Qu et al. TextShield-R1: Reinforced Reasoning for Tampered Text Detection. AAAI, 2026.
Reviewer_e1UaOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how forensic evidence is distributed across the layers of vision transformers and proposes ForenLens, an expert-guided query framework for generalizable image forgery detection. The method uses four expert-guided queries and one free query to extract evidence from selected transformer layers, followed by image-conditioned cross-layer routing to combine the resulting representations. During training, the expert queries are guided by forensic priors derived from Noiseprint++, JPEG Ghost, boundary-transition operators, and SD-VAE reconstruction errors; these auxiliary guidance maps are not required at inference time. The paper also introduces LayerProbe to analyze the layer-wise distribution of forensic information. Experiments across multiple benchmarks and backbones report an average AP of 93.2%, with improvements over several comparison methods using the same SigLIP2 backbone. The paper further evaluates robustness to JPEG recompression and Gaussian blur and reports relatively low additional inference FLOPs. The central claim is that expert-guided queries and adaptive cross-layer evidence aggregation improve generalization by extracting complementary forensic cues while reducing reliance on dataset-specific shortcuts.

Critical strengths

1.The paper addresses an important challenge in AI-generated image and image forgery detection: detectors can overfit to dataset-specific cues and fail under distribution shifts. ForenLens combines expert-guided queries with cross-layer feature aggregation in a coherent framework, while keeping the auxiliary forensic guidance out of the inference pipeline. 2.The paper evaluates multiple backbones and benchmarks and includes comparisons against relevant methods using the same SigLIP2 backbone. The reported average AP of 93.2% and the improvements over the compared methods are encouraging.

Critical weaknesses
  1. The paper lacks controlled comparisons with single-layer features, simple feature concatenation, static layer weighting, and a parameter-matched model without dynamic routing. Thus, the benefit of the proposed routing mechanism remains uncertain.
  2. Ablations using shuffled or mismatched guidance, removing individual priors, and removing expert supervision are needed to determine whether the priors provide meaningful forensic information beyond generic regularization.
  3. Evaluation across benchmarks and under JPEG recompression or Gaussian blur is insufficient to establish robust cross-domain generalization. Generator-disjoint testing, near-duplicate checks, and per-domain performance would strengthen the claims.
  4. The experiments appear to use a single random seed, with no variability estimates. The paper should report multi-seed results and account for guidance-map preprocessing, storage, training cost, and inference latency.
Questions for authors
  1. How does cross-layer routing compare with single-layer features, simple concatenation, static weighting, and a parameter-matched model without dynamic routing?
  2. Do the priors provide distinct forensic information? Could the authors test shuffled or randomized guidance and remove individual priors?
  3. Were generators and image sources strictly separated between training and testing? Could the authors report per-domain results and clarify how data overlap was prevented?
2 comments

We thank you for your constructive comments and provide answers to address each point below.

W1&Q1 (Routing and readout baselines)

We add 4-layer concatenation (Concat-Pooled + Linear, from Pooled in Table 3) and parameter-matched MLP controls (Concat-Pooled + MLP-PM and Queries + Concat MLP-PM) to the comparison. The single-layer (Top-1) and Pooled baselines are provided in Table 3 and Appendix Table 20, and Uniform/Static routing in Table 20.

As shown, the parameter-matched MLPs still lag behind the full query readout. ForenLens consistently outperforms single-layer features, feature concatenation, parameter-matched MLPs, and static/uniform layer weighting on both backbones.

Variant (K=4K=4 unless noted)Dynamic RoutingParams (CLIP / SigLIP2)CLIP APCLIP AUCSigLIP2 APSigLIP2 AUC
Single-Layer (Top-1)×\times8.42M / 10.65M79.580.981.882.6
Concat-Pooled + Linear×\times0.02M / 0.02M79.581.588.989.6
Concat-Pooled + MLP-PM×\times8.42M / 10.66M81.081.090.590.3
Queries + Concat MLP-PM✓\checkmark8.42M / 10.65M79.977.292.892.6
Query + Uniform×\times8.42M / 10.65M82.882.992.392.6
Query + Static×\times8.42M / 10.65M83.283.091.992.3
ForenLens (Query + Adaptive)✓\checkmark8.42M / 10.65M83.583.893.293.3

W2&Q2 (Expert guidance ablations)

We test mismatched guidance by randomly permuting guidance maps across images within each training minibatch, which keeps the spatial distribution of the original guidance maps intact while breaking their alignment with the input image. We compare this with removing spatial guidance (Table 3 and Appendix Table 20), removing expert supervision, and removing individual queries (Table 2 and Appendix Table 19).

VariantCLIP APCLIP AUCSigLIP2 APSigLIP2 AUC
Shuffled / Mismatched Guidance79.278.489.990.2
w/o L_guide80.579.490.591.1
5 Free Queries (no expert supervision)80.178.790.991.2
w/o Noiseprint++80.980.592.693.2
w/o JPEG Ghost80.679.791.992.2
w/o Boundary Transition80.579.492.092.6
w/o SD-VAE Reconstruction82.382.892.793.1
w/o Free Query81.480.892.092.0
ForenLens (Full)83.583.893.293.3

Replacing aligned guidance with shuffled maps degrades detection performance. Moreover, removing any single query or removing expert supervision altogether consistently lowers performance on both backbones, confirming the benefit of aligned guidance and combining the queries.

W3&Q3 (Data separation and per-domain results)

  1. We follow the official OpenMMSec (ICML'26) training, validation, and OOD test splits. The OOD test split contains 72 unseen forgery subtypes across four domains.
  2. The external benchmarks also come from independent datasets. The test generators and source images are disjoint from those used for training and validation. We also checked for exact and near-duplicate images across these splits.
  3. Per-domain results on OpenMMSec are provided in Appendix E.1 (Tables 10–13).

W4 (Multi-seed stability and overhead)

  1. We report the mean and standard deviation over three random seeds (42, 43, and 44), showing stable performance across runs.
Backbone (3 seeds)8-Group AP8-Group AUC8-Group Acc8-Group F1
ForenLens (CLIP)83.4 ± 0.583.6 ± 0.477.8 ± 0.365.6 ± 0.5
ForenLens (SigLIP2)93.2 ± 0.193.2 ± 0.286.8 ± 0.384.5 ± 0.2
  1. Regarding computational and storage overhead:
    1. Offline guidance extraction for Noiseprint++, JPEG Ghost, and SD-VAE is performed only once before training and takes 89.5 ms per image. Preprocessing 1,000 images produces a 4.25 MB cache. Boundary transition is computed online from the augmented image.
    2. Training cost and inference latency are detailed in Appendix D.3 (Tables 8–9).
    3. At inference, no guidance maps need to be computed or used.
AuthorsOpenReview ↗

We look forward to your responses

Dear Reviewer e1Ua,

We have provided point-wise clarification and experiments (upon your request) for all your concerns. We look forward to your reply and hope that our responses could address your concerns. Your time and effort is very much appreciated.

Best Regards

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the AC and reviewers for their feedback and are glad they recognized our main findings:

  • extracting forgery cues already in pretrained visual encoders is convincing (aBB2);
  • the 14-backbone study gives a systematic characterization of layer-wise discriminability and complementarity (xWc2);
  • ForenLens unifies expert-guided queries and cross-layer readout in a coherent framework without inference-time guidance (e1Ua, xWc2);
  • multi- and same-backbone evaluations support its effectiveness (e1Ua, xWc2) at reduced training cost (xWc2).

Below we summarize our clarifications and new experiments (full tables in individual replies).

1. Where gains come from on a fixed frozen backbone (e1Ua, xWc2, aBB2). With the frozen backbone, training data, and splits fixed, we isolated each readout part on CLIP and SigLIP2:

Same-Backbone Control (K=4 unless noted)CLIP APSigLIP2 AP
Single selected layer (K=1 query)79.581.8
Concat-pooled + capacity-matched MLP81.090.5
Top-4 individual layers (Rank-4)79.591.6
Queries w/o spatial guidance80.590.5
Queries w/ shuffled guidance maps79.289.9
ForenLens (full)83.593.2

A capacity-matched MLP on concat-pooled layers still lags, showing that preserving layer identity via query retrieval and adaptive routing extracts evidence flat pooling washes out (uniform/static weights also trail). Rank-4 drops CLIP to single-layer AP, as ranking clusters in adjacent middle blocks and misses deeper cues. On the same SigLIP2 checkpoint and splits, ForenLens also beats frozen linear probing (UnivFD, 87.1 AP) and adaptation baselines (SICA, 88.6; Effort, 89.9).

2. Shallow and deep layers add complementary forensic cues, not global semantics (xWc2, aBB2). On unseen data, adjacent ViT blocks are highly correlated with minimal fusion gain, whereas greedily paired shallow and deep blocks are far less correlated and correct over half of each other's errors. Crucially, this is not driven by scene semantics: on unseen regional manipulations (face editing/swapping/reenactment and non-generative IMDL) with nearly identical real/fake content, 4-layer fusion boosts OOD AUC by 16–29 points over the single-layer baseline. Since neighboring blocks are similar, single-domain probes pick different indices yet nearly always pair early/middle with deep layers, keeping OOD gains.

3. Expert priors guide where queries look rather than deciding real vs. fake (e1Ua, xWc2, aBB2). Our four experts use established cues (noise residuals, JPEG traces, boundary transitions, reconstruction error) for training-only supervision (spatial attention, plus a scalar target for SD-VAE), leaving classification to the readout. Batch-permuted maps keep spatial statistics while breaking alignment and drop AP below no spatial guidance (table above); removing any query hurts both backbones. On benign structures in real images, priors only suggest candidate regions to inspect, with guided queries matching target maps and the free query attending elsewhere.

4. One fixed detector transfers across domains and recent generators without readaptation (e1Ua, xWc2, aBB2). Trained once on OpenMMSec source data, one fixed detector is tested unchanged on 72 unseen OpenMMSec subtypes and external benchmarks:

  • Balanced cross-domain generalization. Grouping all AIGC sets into one 25% share with Deepfake, IMDL, and Document (25% each) removes the 8-group AIGC tilt; ForenLens leads all four metrics, beating SICA in AP (82.8 vs. 78.6) and accuracy (82.9% vs. 81.4%).
  • From-scratch synthesis and recent generators. When boundary cues are less informative on from-scratch synthesis (e.g., AIGI-S, Chameleon), our readout down-weights the boundary query toward compression and reconstruction cues. Beyond 9 modern generators in AIGI-Now (95.6/89.7 AP on its two tracks), the fixed detector leads all 10 recent Rapidata generator subsets (several post-submission, e.g., Seedream 5.0 Pro, GPT Image 2.5), raising mean fake accuracy from 53.1% (SICA) to 67.0% without readaptation.

5. Stability, layer budget, and compute overhead (e1Ua, aBB2). Over 3 seeds, standard deviations are ≤0.5 AP, and sweeping 1–10 blocks saturates beyond K=4, supporting one shared 4-layer budget across all 14 encoders. Like UnivFD, RINE, Effort, SICA, and PGC, ForenLens trains once per encoder; offline caching and training-only guidance keep preprocessing light and inference operator-free. Our readout adds <3% forward FLOPs, and freezing the backbone halves training compute vs. same-backbone adaptation methods, with lower single-image latency than those methods on SigLIP2.

We will add these ablations, balanced cross-domain results, and recent-generator tests to the revision, and clarify our zero-readaptation OOD setting. Thank you again.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.